← sources

GLM-5.3-Flash on our H200s does most of my coding now while GPT-6.1 Sol orchestrates (OpenCode setup)

source: reddit (reddit:LocalLLaMA)published 2026-11-01 09:49ingested 2026-11-02 02:41

retained text

I had been trying this setup for a couple of weeks and it  working better than I expected. Recently, I deployed [GLM-5.3-Flash](https://huggingface.co/zai-org/GLM-5.3-Flash) (MoE, 320B total, 18B active) on one of our local H200 servers (4 GPUs) with a 512K context window with the aim of how can I use it as my main coding workhorse and  
cut down my subscription and API costs 🙈.

I was very quickly able to realize, I could not switch to it for everything but pairing it with GPT-6.1 Sol works really well with Sol being the the orchestrator and GLM being the subagents model that explores, writes and implements the code.

Honestly, I would have preferred doing this in Claude Code or Codex since I am used to them but neither lets a session or its subagents use models from different providers in a single session. This is where I found OpenCode flexible that lets agent have its own `provider/model`, so GLM goes as sub-agent models along with the added benefit of chatGPT subscription being available too (Allows you to use Sol too!)/

By the way, nothing here is gml specific any model you can serve works as the workhorse. I have also written a blogpost about it: [https://aayushgarg.dev/posts/2026-11-01-gpt-local-model-opencode/](https://aayushgarg.dev/posts/2026-11-01-gpt-local-model-opencode/)

I would love to know about other people setup, how they are using open-source models either individually or pairing them with opus/sol/astra.

document fetched 2026-11-02 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1xkz4lf/glm53flash_on_our_h200s_does_most_of_my_coding